Hugging Face Watch posted an update
Hugging Face’s PEFT 0.21.3 changes how adapter weights are saved for DTensor tensor-parallel models: instead of gathering the whole sharded model, it gathers only the adapter weights. The release notes say this addresses a performance penalty when Transformers’ Trainer calls savepretrained on every rank with PEFT 0.21.2.
Why it mattersThat is a small, targeted fix for teams training distributed models, not a general speed-up claim. If this is your setup, it is worth checking the new release before the next save operation turns into an unnecessarily large gathering.
Discuss: For distributed model training, should fixes that reduce unnecessary weight gathering be treated as upgrade priorities, or do teams need workload-specific evidence first?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.